Thanks for contributing these! Putting into draft for now — we'll review properly after the hackathon wraps up.

A few things we'll want to address before merging:

1. Quality pass needed — Both skills are heavy on pseudocode and aspirational workflow descriptions. We'd want to tighten them up with tested, actionable instructions that actually work end-to-end.
2. Terminal artifacts — Both files have a leaked fence string at the end that needs to be removed.
3. Dependency references — A few skill names in the autonomous-researcher metadata are wrong (hermes-agent-spawning → hermes-agent, evaluating-llms-harness → lm-evaluation-harness).
4. Bundled vs Hub — These are fairly specialized (Tinker API key required, ML research niche). May be better as Skills Hub entries rather than bundled with every install.

Will circle back after the hackathon!

Orb Code Review (GLM 5.1 on Orb Cloud)

PR #760 Review: feat(skills): add autonomous research and tinker GPU training capabilities

Reviewed both new SKILL.md files: skills/mlops/tinker/SKILL.md and skills/research/autonomous-researcher/SKILL.md.

🐛 Bugs

1. Wrong module reference in Supervised Learning example (tinker SKILL.md)

trainer.optim_step(types.AdamParams(learning_rate=1e-4)).result()

types is never imported — should be tinker.AdamParams to match every other example in the file.

2. Undefined data variable in code examples
Multiple code blocks reference data or batch without defining them (Quick Start, Supervised Learning, train_chunk.py). Users copying these examples will hit NameError. Consider adding a placeholder like data = load_batch() or noting it with a comment.

3. Missing Mode D example invocation
The Input Modes section defines Modes A–E, but the Example Invocations section jumps from Mode C to Mode E. Mode D ("User-Provided Everything") has no corresponding example.

🔒 Security

4. Shell injection in self-chaining pattern

f"hermes chat -q 'Continue training with checkpoint {decision["checkpoint_path"]}'"

The checkpoint_path from the API response is interpolated directly into a shell command without sanitization. If the path is manipulated or contains shell metacharacters, this is a command injection vector. Consider using shlex.quote() or passing the value as a separate argument.

5. git add . && git push without review
The publishing workflow includes git add . which could accidentally commit secrets, large binaries, or other unintended files. Consider being more explicit about what gets committed.

📋 Content / Quality

6. Trailing HERMES_EOF artifacts in both files
Both SKILL.md files end with garbage text:

HERMES_EOF; __hermes_rc=$?; printf '__HERMES_FENCE_a9f7b3__'; exit $__hermes_rc

This appears to be a session fence marker that was accidentally included in the file content. It should be removed, and both files should end with a proper newline.

7. Massive duplication between the two files
Phases 4–8 (Training, Evaluation, Iteration, Writing, Publishing), the Skill Routing Table, Example Invocations, Quick Reference Commands, and Troubleshooting sections are nearly identical in both files. This creates a significant maintenance burden — any fix or update must be applied in two places. Consider having the autonomous-researcher skill reference the tinker skill for training-specific details rather than duplicating them.

8. tinker checkpoint download is undocumented
The evaluation workflow references tinker checkpoint download tinker://model_id/final as a CLI command, but this command doesn't appear in the Tinker skill's Core API or Setup sections. Is this a real CLI command or should it be a Python API call?

Summary

The overall structure and concept of both skills is solid — the 8-phase research pipeline and self-chaining session pattern are well-designed. However, the code examples contain bugs that would confuse users, there's a security concern with shell interpolation, and the trailing HERMES_EOF artifacts need to be cleaned up before merging. The duplication between files is the largest maintenance concern.

Recommendation: Request changes — address the HERMES_EOF artifacts, the types.AdamParams bug, undefined variables in examples, and the shell injection risk.

Charon Code Review — PR #760

Two new skills: Tinker (GPU fine-tuning via Thinking Machines Lab API) and Autonomous Researcher (end-to-end ML research pipeline). Both are SKILL.md files — primarily documentation/instructional content. Structural review below.

🔴 Critical — ml-paper-writing skill listed as required but does not exist

autonomous-researcher/SKILL.md metadata:

requires:
  - ml-paper-writing   # ← listed here

ml-paper-writing is not in the available skills registry. Phase 7 (Writing) routes to this non-existent skill. The skill invocation will fail at runtime. Either remove it from requires and the routing table, or create the actual skill first.

🔴 Critical — Tinker skill has a Python syntax error in RL training example

tinker/SKILL.md:

for iteration in range(15):
    sampler = trainer.save_weights_and_get_sampling_client(f"iter_{iteration}")

The opening f" is missing before iter_. This causes a SyntaxError if copied. Fix: trainer.save_weights_and_get_sampling_client(name=f"iter_{iteration}").

🟡 Warning — lm-eval results access pattern is wrong

autonomous-researcher/SKILL.md, Phase 6:

if results["mmlu"] < baseline["mmlu"]:
    issues.append("MMLU degraded")

The lm-evaluation-harness returns nested results like {"mmlu": {"mmlu": {"acc": 0.72}}} or {"mmlu": {"acc": 0.72}}. Not flat {"mmlu": 0.72}. This raises KeyError or compares dict to float. Fix: results.get("mmlu", {}).get("acc", 0).

🟡 Warning — data variable is undefined in train_chunk example

tinker/SKILL.md:

for step in range(args.steps):
    result = trainer.forward_backward(data, "cross_entropy")  # data never defined
    losses.append(result.loss)

data is never loaded, defined, or constructed. CHUNK_SIZE, CHECKPOINT_EVERY, load_state(), save_state() are also undefined. The example crashes immediately. Add a data loading stub or make the incompleteness explicit.

🟡 Warning — mkdir should not be background

autonomous-researcher/SKILL.md, Phase 3:

terminal(command="mkdir -p ...", background=True)

mkdir is synchronous and instant. Backgrounding it provides no benefit. Should be background=False.

🟡 Warning — initialize_project() is pseudocode, not runnable

autonomous-researcher/SKILL.md, Phase 0: The function references undefined determine_mode(), calls write_file() without import, and mixes Python syntax with markdown prose. Label clearly as pseudocode/description, not executable code.

🟢 Nit — Routing table lists 8 phases but ml-paper-writing doesn't appear in it

The routing table covers Phases 1-8, but ml-paper-writing (Phase 7 Writing) is missing from the table while listed in requires. Inconsistency between metadata and routing table.

🟢 Nit — compute_rewards and compute_advantages are undefined stubs

tinker/SKILL.md, RL Training Pattern:

rewards = compute_rewards(samples, ground_truth)   # undefined
advantages = compute_advantages(rewards)            # undefined

These functions are never defined. Either implement minimal stubs or point to trl / openrl as the source.

✅ What is solid

• Tinker API documentation is comprehensive — architecture diagram, model table, loss function table, and core API examples are all clear and correct
• Self-chaining session concept is the right architectural pattern for long-running training across Hermes sessions
• Autonomous Researcher pipeline is well-designed — 8-phase pipeline with clear decision points and mode routing (A-E) is sound
• Checkpoint patterns (weights-only vs full state) are clearly distinguished
• Iteration decision tree is a useful operational guide
• Skill routing table correctly maps phases to primary and secondary skills

Verdict: Request Changes — two criticals (non-existent skill reference, Python syntax error) must be fixed. The lm-eval access pattern and undefined variables are functional bugs that cause runtime failures.